Real-time monitoring of long-overhang boring conditions using multi-sensor fusion and time–frequency deep learning

Machining condition monitoring in long-overhang boring operations faces substantial challenges, including severe noise contamination, signal degradation, and feature instability arising from tool deflection, vibration coupling, and dynamic changes in cutting load. Traditional methods mainly rely on single-sensor signals or construct offline classification models under specific machining conditions, lacking cross-parameter adaptability and real-time identification capabilities, which makes them difficult to meet practical machining needs. To address these issues, this study proposes a machining monitoring architecture that combines multi-sensor fusion with deep learning. The system integrates three-axis vibration and two-phase current signals from the spindle, extracts time-frequency features using a continuous wavelet transform, and uses static machining parameters such as spindle speed, feed rate, and cutting depth as auxiliary labels to construct a multi-modal fusion model. The model performs online inference on a Computer Numerical Control lathe, enabling real-time identification of three operating states: normal, incipient abnormal, and severely abnormal. Experimental results demonstrate that the proposed method achieves classification accuracy exceeding 95% across varied machining conditions while meeting the latency requirements for online monitoring. The findings confirm that combining multi-sensor fusion with a time-frequency deep model can effectively improve the monitoring stability and deployability of long-overhang boring, demonstrating its potential for smart manufacturing applications.